A Lagrangian dual relaxation approach to ML MIMO detection: Reinterpreting regularized lattice decoding

This paper describes a new approximate maximum-likelihood (ML) MIMO detection approach by studying a Lagrangian dual relaxation (LDR) of ML. Unlike many existing relaxed ML methods, the proposed LDR employs a discrete domain for the problem formulation. We find that the proposed LDR exhibits an intriguing relationship to the lattice decoders (LDs) and the lattice reduction aided (LRA) detectors, both of which have caught much attention recently. Specifically, regularization in LDs, which was proposed to mitigate out-of-bounds symbol effects, can alternatively be interpreted as a way to constrain the symbol decision within bounds in a Lagrangian sense. We handle the LDR problem by using a projected subgradient method. The resultant method may physically be viewed as an adaptive regularization control in which a sequence of LDs are involved. Based on this newly developed insight, we propose two additional iterative LDR-based detectors using LRA decision-feedback (DF) and "lazy" DF. By simulation results, we show that the LDR LRA-DF and lazy-DF detectors yield better symbol error rate performance than the MMSE-regularized LRA-DF and DF detectors, respectively, where the SNR gaps can be more than 3dB.